LLM RAG Patterns Cheat Sheet

Retrieval-Augmented Generation patterns — chunking strategies, embedding, vector search, reranking, hybrid search, and RAG evaluation metrics.

Last Updated: July 15, 2025

RAG Architecture Pipeline

1. Ingest

Load documents — PDF, HTML, Markdown, APIs

2. Chunk

Split into retrievable units — fixed-size, semantic, recursive

3. Embed

Convert chunks to vectors — OpenAI, Cohere, sentence-transformers

4. Store

Index in vector DB — Pinecone, Weaviate, Chroma, Qdrant

5. Retrieve

Query → embed → similarity search (cosine, dot product, Euclidean)

6. Augment

Inject top-k chunks + user query into LLM prompt

Chunking Strategies

Strategy	How It Works	Best For
Fixed-Size	Split by N tokens with overlap	General purpose, fast
Recursive	Split by separators: \n\n → \n → . → space	Markdown/docs
Semantic	Split at natural topic boundaries (embeddings)	Long-form content
Sentence Window	Retrieve sentence, expand context window	Precision retrieval

Retrieval Techniques

Technique	Description
Dense Retrieval	Embed query and chunks — semantic similarity
Sparse Retrieval	BM25/TF-IDF keyword matching — exact matches
Hybrid Search	Combine dense + sparse → weighted fusion (RRF)
Multi-Query	Generate variant queries → retrieve → deduplicate union
Parent Document	Retrieve small chunks, return full parent document

Reranking

Method	Description
Cross-Encoder	Score (query, chunk) pairs — Cohere Rerank, BGE-reranker
LLM-as-Reranker	GPT-4 evaluates relevance of each chunk
Reciprocal Rank Fusion	Merge multiple ranked lists without scores

Vector Databases Compared

DB	Self-Hosted	Filtering	Best For
Pinecone	No (cloud)	Metadata + namespaces	Production, scale
Weaviate	Yes	GraphQL + hybrid search	Multimodal, hybrid
Chroma	Yes	Metadata filtering	Prototyping, local dev
Qdrant	Yes	Payload filtering	Performance, Rust-based
Milvus	Yes	Scalar + vector	Billion-scale, GPU indexing

Pro Tip: Chunk size is the most impactful RAG decision — too small loses context, too large dilutes relevance. Start at 512 tokens with 64-token overlap and tune from there.